--------------------------------------------------------------------
“神經網絡”這個詞實際是來自於生物學，而我們所指的神經網絡正確的名稱應該是“人工神經網絡（ANNs）”。在本文，我會同時使用這兩個互換的術語。
壹個真正的神經網絡是由數個至數十億個被稱為神經元的細胞（組成我們大腦的微小細胞）所組成，它們以不同方式連接而型成網絡。人工神經網絡就是嘗試模擬這種生物學上的體系結構及其操作。在這裏有壹個難題：我們對生物學上的神經網絡知道的不多！因此，不同類型之間的神經網絡體系結構有很大的不同，我們所知道的只是神經元基本的結構。
The neuron
----------------------------------------------------------------------
雖然已經確認在我們的大腦中有大約50至500種不同的神經元，但它們大部份都是基於基本神經元的特別細胞。基本神經元包含有synapses、soma、axon及dendrites。Synapses負責神經元之間的連接，它們不是直接物理上連接的，而是它們之間有壹個很小的空隙允許電子訊號從壹個神經元跳到另壹個神經元。然後這些電子訊號會交給soma處理及以其內部電子訊號將處理結果傳遞給axon。而axon會將這些訊號分發給dendrites。最後，dendrites帶著這些訊號再交給其它的synapses，再繼續下壹個循環。
如同生物學上的基本神經元，人工的神經網絡也有基本的神經元。每個神經元有特定數量的輸入，也會為每個神經元設定權重（weight）。權重是對所輸入的資料的重要性的壹個指標。然後，神經元會計算出權重合計值（net value），而權重合計值就是將所有輸入乘以它們的權重的合計。每個神經元都有它們各自的臨界值（threshold），而當權重合計值大於臨界值時，神經元會輸出1。相反，則輸出0。最後，輸出會被傳送給與該神經元連接的其它神經元繼續剩余的計算。
Learning
----------------------------------------------------------------------
正如上述所寫，問題的核心是權重及臨界值是該如何設定的呢？世界上有很多不同的訓練方式，就如網絡類型壹樣多。但有些比較出名的包括back-propagation, delta rule及Kohonen訓練模式。
由於結構體系的不同，訓練的規則也不相同，但大部份的規則可以被分為二大類別 - 監管的及非監管的。監管方式的訓練規則需要“教師”告訴他們特定的輸入應該作出怎樣的輸出。然後訓練規則會調整所有需要的權重值（這是網絡中是非常復雜的），而整個過程會重頭開始直至數據可以被網絡正確的分析出來。監管方式的訓練模式包括有back-propagation及delta rule。非監管方式的規則無需教師，因為他們所產生的輸出會被進壹步評估。
Architecture
----------------------------------------------------------------------
在神經網絡中，遵守明確的規則壹詞是最“模糊不清”的。因為有太多不同種類的網絡，由簡單的布爾網絡（Perceptrons），至復雜的自我調整網絡（Kohonen），至熱動態性網絡模型（Boltzmann machines）！而這些，都遵守壹個網絡體系結構的標準。
壹個網絡包括有多個神經元“層”，輸入層、隱蔽層及輸出層。輸入層負責接收輸入及分發到隱蔽層（因為用戶看不見這些層，所以見做隱蔽層）。這些隱蔽層負責所需的計算及輸出結果給輸出層，而用戶則可以看到最終結果。現在，為免混淆，不會在這裏更深入的探討體系結構這壹話題。對於不同神經網絡的更多詳細資料可以看Generation5 essays
盡管我們討論過神經元、訓練及體系結構，但我們還不清楚神經網絡實際做些什麽。
The Function of ANNs
----------------------------------------------------------------------
神經網絡被設計為與圖案壹起工作 - 它們可以被分為分類式或聯想式。分類式網絡可以接受壹組數，然後將其分類。例如ONR程序接受壹個數字的影象而輸出這個數字。或者PPDA32程序接受壹個坐標而將它分類成A類或B類（類別是由所提供的訓練決定的）。更多實際用途可以看Applications in the Military中的軍事雷達，該雷達可以分別出車輛或樹。
聯想模式接受壹組數而輸出另壹組。例如HIR程序接受壹個‘臟’圖像而輸出壹個它所學過而最接近的壹個圖像。聯想模式更可應用於復雜的應用程序，如簽名、面部、指紋識別等。
The Ups and Downs of Neural Networks
----------------------------------------------------------------------
神經網絡在這個領域中有很多優點，使得它越來越流行。它在類型分類/識別方面非常出色。神經網絡可以處理例外及不正常的輸入數據，這對於很多系統都很重要（例如雷達及聲波定位系統）。很多神經網絡都是模仿生物神經網絡的，即是他們仿照大腦的運作方式工作。神經網絡也得助於神經系統科學的發展，使它可以像人類壹樣準確地辨別物件而有電腦的速度！前途是光明的，但現在...
是的，神經網絡也有些不好的地方。這通常都是因為缺乏足夠強大的硬件。神經網絡的力量源自於以並行方式處理資訊，即是同時處理多項數據。因此，要壹個串行的機器模擬並行處理是非常耗時的。
神經網絡的另壹個問題是對某壹個問題構建網絡所定義的條件不足 - 有太多因素需要考慮：訓練的算法、體系結構、每層的神經元個數、有多少層、數據的表現等，還有其它更多因素。因此，隨著時間越來越重要，大部份公司不可能負擔重復的開發神經網絡去有效地解決問題。
NN 神經網絡，Neural Network
ANNs 人工神經網絡，Artificial Neural Networks
neurons 神經元
synapses 神經鍵
self-organizing networks 自我調整網絡
networks modelling thermodynamic properties 熱動態性網絡模型
英文翻譯
Introduction
----------------------------------------------------------------------
Neural network is a new technology in the field of fashion vocabulary. Many people have heard of the word, but few people really understand what it is. The purpose of this paper is to introduce all the basic neural network functions, including its general structure, related terms, types and applications.
"Neural network" actually came from biology, and neural networks we refer the correct name should be "Artificial Neural Networks (ANNs)". In this article, I will also use the two interchangeable terms.
A real neural network is a few to a few billion cells called neurons (composed of tiny cells in our brains) are composed of, they are different ways to connect and type into the network. Artificial neural network is trying to model this biological system structure and its operation. There is a problem here: we biological neural networks do not know much! Thus, between different types of neural network architecture is very different, we know only the basic structure of neurons.
The neuron
----------------------------------------------------------------------
While already recognized in our brain, about 50 to 500 kinds of different neurons, but most of them are based on special cells in the basic neuron. Contains the basic neural synapses, soma, axon and dendrites. Synapses between neurons responsible for the connection, they are not directly physically connected, but they have a very small gap between to allow electronic signals from one neuron to another neuron. Then the electrical signals to the soma will be an internal electronic signal processing and its processing result will pass axon. The axon of these signals will be distributed to dendrites. Finally, dendrites with these signals and then to the other synapses, and then continue to the next cycle.
As a basic biological neurons, artificial neural networks have basic neurons. Each neuron has a specific number of inputs, will be set for each neuron weight (weight). Weight is the importance of the information entered an indicator. Then, neurons calculates the weight of the total value (net value), while the total weight of all the input value is multiplied by the total of their weights. Each neuron has their own threshold (threshold), while the power is greater than the critical value of the total value of weight, the neuron will output 1. On the contrary, the output 0. Finally, the output can be transmitted to the neuronal connections with other neurons to the remaining calculations.
Learning
---------------------------------------------------------------------- As written above, at issue is the critical value of the weight and how to set it? The world has many different training methods, as much as the network type. But some well-known, including back-propagation, delta rule and Kohonen training mode.
Because of different structural systems, training is not the same rules, but most of the rules can be divided into two broad categories - regulatory and non-regulated. Supervising the training rules need to be "teachers" tell them how a particular input to the output should be. Then the training rule to adjust the weight of all the needs of value (this is a very complex network), and the whole process would start again until the correct data can be analyzed by the network. Regulatory approach of the training model includes back-propagation and the delta rule. The rules of non-regulatory approach without teachers, because they produce the output will be further evaluated.
Architecture
----------------------------------------------------------------------
In the neural network, comply with the rules clear word is the most "obscure" the. Because there are too many different types of networks, from simple Boolean networks (Perceptrons), to the complex network of self-adjustment (Kohonen), to the thermal dynamic network model (Boltzmann machines)! These have to comply with the standards of a network architecture.
A network including multiple neurons, "layer", the input layer, hidden layer and output layer. Input layer to receive input and distribute to the hidden layer (because the user can not see the layers, so do see the hidden layer). The hidden layer is responsible for the necessary calculations and output to the output layer, the user can see the final result. Now, to avoid confusion, would not be here more in-depth study architecture talking about it. Different neural networks for more detailed information can be read Generation5 essays, including a multiple neural network "layer", the input layer, hidden layer and output layer. Input layer to receive input and distribute to the hidden layer (because the user can not see the layers, so do see the hidden layer). The hidden layer is responsible for the necessary calculations and output to the output layer, the user can see the final result. Now, to avoid confusion, would not be here more in-depth study architecture talking about it. Different neural networks for more detailed information can be seen Generation5 essays.
Although we discussed the neurons, training and architecture, but we do not know what the actual neural network.
The Function of ANNs
---------------------------------------------------------------------- Neural networks are designed to work with patterns - they can be divided into two categories-type or association type. Category-type network can accept a few, and then classified. For example, ONR program accepts a number of the image and the output figure. Or PPDA32 program accepts a coordinate and to classify it as Class A or B (type of training provided by the decision). More practical use can be seen Applications in the Military in the military radars, the radar could pick out a vehicle or tree.
Lenovo model to accept a group of numbers and the output of another group. HIR procedures such as acceptance of a 'dirty' image and the output of a learned and the closest it an image. Lenovo model also can be used in complex applications such as signature, face, fingerprint recognition.
The Ups and Downs of Neural Networks
---------------------------------------------------------------------- Neural network in this area has many advantages, making it more popular. It is in the type classification / recognition is very good. Neural networks can handle the exception and not the normal input data, which are important for many systems (such as radar and sonar systems). Many neural networks are mimic biological neural networks, that is their mode of operation modeled on the work of the brain. Neural networks also have to help the development of neuroscience, it can, like humans, accurate identification of objects and the speed of computers! The future is bright, but now ...
Yes, the neural network are also some bad points. This is usually because of lack of sufficiently powerful hardware. Power derived from the neural network to process information in parallel, that is, a number of data simultaneously. Therefore, to simulate a serial parallel processing machines is very time-consuming.
Another problem with neural networks is a problem in building a network of defined conditions are not - there are too many factors to consider: training algorithms, architecture, number of neurons in each layer, the number of layers, data show, etc. There are other additional factors. Therefore, more and more important over time, most companies can not afford to repeat the development of neural network to effectively solve the problem.
不知道是不是 我隨便找的